Fast Gaussian blur implementation that runs entirely on the CPU using an Infinite‑Impulse‑Response (IIR) filter. The code is written in C++ and comes in two flavours:
- Scalar ‑ portable, works on every Android CPU
- NEON ‑ hand‑optimized SIMD kernels that are up to 3× faster on ARMv8 devices
for comparison fp16 version, box blur and gpu blur also added
For real use bitmaps may be pre-downscaled
Grab the sample application:
The algorithm is based on the Deriche formulation of the Gaussian blur. Instead of convolving with a large kernel (O(r²)), it uses recursive filters that achieve the same visual result in O(N) time, independent of the radius. The two‑pass (horizontal + vertical) implementation keeps the memory footprint low. When NEON is available we process four pixels per register and interleave loads/stores to maximise throughput.